Following (DuyBui, 2004; Sehaba et al., 2007), three .... type of emotion, the solidarity decreases (Figure 3). ... policeman and the Burglar triggered emotions. 1.
Modeling the Dynamics of Non-Player Characters' Social Relations in Video Games
Magalie Ochs, Nicolas Sabouret and Vincent Corruble
LIP6 - Université Pierre & Marie Curie, France
Abstract
player.
Building credible Non-Playing Characters (NPCs) in games requires not only to enhance the graphic animation but also the behavioral model. This paper tackles the problem of the dynamics of NPCs social relations depending on their emotional interactions. First, we discuss the need for a dynamic model of social relations. Then, we present our model of social relations for NPCs and we give a qualitative model of the inuence of emotions on social relations. We describe the implementation of this model and we briey illustrate its features on a simple scene.
In many games, this is done to some extent
through some relatively simple means. an alignment (good vs.
evil, loyal vs.
For example, chaotic) can
constrain available actions and induce specic types of reactions from NPCs (for example in Baldur's Gate). In other games, the Player Character's actions change its reputation in the virtual world, aecting indirectly the attitude of NPCs such as guards or merchants, up to a point where, respectively, after enough misdeeds, he/she could be attacked at rst sight, or be oered less favorable deals.
Though interesting, these means re-
mains relatively crude from the cognitive science standpoint. Various results from psychology can be used to elaborate a ner representation of social relations. However, most of the models available in the literature are
Introduction
static in nature. In this paper we propose a model of
One of the main drives for the Game AI community has
social relations between synthetic characters that is dy-
been to develop synthetic characters, including Non-
namic social relations will evolve over time and
Player Characters (NPCs), that are more credible and
connected to an emotional model. The methodology is
engaging. Beyond the physical dimension (graphics, an-
to infer as much as possible from the psychology liter-
imation, etc.), the emotional dimension of NPCs is usu-
ature so that the implementation of the model requires
ally seen as key to preserve the immersive quality of a
only a small number of parameters to be set. The ap-
virtual world. Though many aspects of the game can
proach can be used to compute the evolution of social
be scripted, the player's actions cannot and therefore
relations over an entire scenario, or also within a single
they (or rather the Player Character's actions) need to
scene such as a dialog.
have visible and believable impacts on the other NPC's actions, through the emotions they generate.
Model-
This paper is organized as follows. cus related work.
First, we dis-
Secondly, we introduce a computa-
ing emotions credibly leads to NPCs who appear more
tional representation of Non-Player Characters' social
responsive, reactive to the player's playstyle.
How-
relations. Thirdly, we describe an emotion-based model
ever, NPC's emotional reactions should follow their own
of the dynamics of NPC's social relations. To conclude,
logic, which is in part character dependent.
we present the implementation of the proposed model
For in-
stance, a character's emotional reaction should vary de-
and test it with a single dialog.
pending on the character's personality. This line of research has been explored in previous work, including in (Sehaba et al., 2007). Another important issue is the impact of social relations on the NPCs emotional reactions. This also is key since in many games, especially role-playing games and adventure games, the player has to impersonate a specic role in a complex story. Social relations are therefore somewhat pre-dened by the scenario, but the evolution of the story and of social relations needs to take into account the choices of the
c 2008, Association for the Advancement of ArCopyright ticial Intelligence (www.aaai.org). All rights reserved.
Related work
Several researchers have proposed to model the social context of virtual characters to increase their believability. For instance, (Rousseau and Hayes-Roth, 1998) suggest a social-psychological model including personality, emotions, and attitudes. The character's behavior is computed based on its personality and attitudes. In (Walker et al., 1997) and (André et al., 2004), the linguistic style of dialog between virtual character is determined according to social variables (social distance and power) (Walker et al., 1997) and user's emotions (An-
dré et al., 2004). (Gratch, 2000) proposes a social layer
2001). Based on this literature, in our model, we con-
that manages communication and biases plan genera-
sider these four social variables to represent a social
tion and execution in accordance with a social context.
relation.
In (Prendinger and Ishizuka, 2001), social lter pro-
Note that social relations are directed and not neces-
grams are proposed to constrain a character's expres-
sarily symetric. For instance, the fact that a character
sion of emotion depending on the social context (rep-
i
resented through social power and social distance), the
that
virtual character's personality, and its emotions. (Bick-
Ishizuka, 2001; Bickmore and Cassell, 2001), the social
more and Cassell, 2001) propose to use the character's
relations are dened from the point of view of a given
social relations with the user to identify the appropriate
virtual character.
likes another character
j
likes
i.
j
does not necessarily mean
In our model, like in (Prendinger and
subject to discuss with the user. As it turns out, the models of social context in existing work are generally
static.
Researchers have mainly
Representation of NPC's social relation
In our dynamic model of social relations, the relation
focused on the impact of social context on virtual char-
of a NPC
acter's behavior.
of
Few researchers propose a model of
i,
i
at time
with a character
t
j,
from the point of view
is formally represented as follows:
the dynamics of social context. (Prendinger et al., 2002) propose functions to model the dynamics of a character's attitude and familiarity toward another character according to the valence of emotions triggered during the interaction, i.e. their positive (e.g. joy) or negative (e.g. fear) aspect. However, this model does not take into account the distinct impacts of dierent types of emotions.
Research shows that emotions of same va-
lence but dierent types (like shame and distress) may have dierent impacts on social relations (Shiota et al., 2004; de Rivera and Grinkis, 1986). In this paper, we go further by proposing a model of the dynamics of character's social relations based on emotions, in checking those of the character and those expressed by its interlocutor.
likingi,j (t) dominancei,j (t) social_relationi,j (t) = f amiliarityi,j (t) solidarityi,j (t)
with:
• likingi,j (t) ∈ [−1, 1] represents the degree of liking i has for j at time t. The closer the value is to −1 (resp. 1), the more i dislikes (resp. likes) j ; • dominancei,j (t) ∈ [−1, 1] corresponds to the power that i thinks it can exert on j at time t. The closer this value is to −1 (resp. 1), the more i feels submissive to (resp. dominant on) j ; • f amiliarityi,j (t) ∈ [0, 1]
Non-Player Characters' Social Relations
In computational models of social characters, social relations are generally represented by a nite set of variables. Each of them characterizes a specic dimension of a social relation between two agents (virtual or human). No consensus exists on the type and number of variables required to model social relations. However, the literature seems to outline four main social vari-
degree of liking
(Isbister, 2006; Rousseau
and Hayes-Roth, 1998; Prendinger and Ishizuka, 2001)
1
one has for another ; the Rousseau
and
dominance
Hayes-Roth,
1998;
(Isbister, 2006;
Prendinger
and
Ishizuka, 2001), i.e. the power that an agent can exert on another agent; the sell, 2001), a.k.a.
solidarity
(Bickmore and Cas-
social distance (Brown and Levin-
son, 1987), is sometimes used (Walker et al., 1997; André et al., 2004), which can be dened as the degree of like-mindedness or having similar behavior dispositions (e.g.
i of the information that i j . A value close to 1 means that i feels familiar with j (in terms of condentiality of the information transmitted). If i feels unfamiliar with j , the value is equal to 0; has transmitted to
Background
ables: the
represents the average de-
gree of condentiality for
similar political membership, family, re-
ligions, profession, gender, etc.) sell, 2001); the
(Bickmore and Cas-
f amiliarity may be used to characterize
• solidarityi,j (t) ∈ [0, 1]
represents the degree of like-
mindedness or having similar behavioral dispositions
i thinks to have with j . The closer the value 1, the more i thinks to share equivalent goals, beliefs, and values. The value 0 represents a situation in which i does not think to share same goals, beliefs or values with j . that
is to
As stated above, social relations are not necessarily symmetric, i.e. :
∃i, j social_relationi,j (t) 6= social_relationj,i (t) The next section presents our emotion-based model of the dynamics of social relation. The Emotion-based Dynamics of Social Relations
the type (private or public) and number of information
Research shows that, during an interaction, one's emo-
exchanged between two agents (Bickmore and Cassell,
tions and those of his interlocutor may lead to a change
1
The term attitude is sometimes used to refer to the degree of liking.
in their social relations (Ortony, 1991; de Rivera and Grinkis, 1986; Shiota et al., 2004). In this paper, we do not focus on the emotion model, which has been widely
studied in the litterature, but rather on the dynamics of the social relations. Thus, we use the well-known OCC model (Ortony, 2002) to represent the NPC's emotional state, which considers the following pairs of emo-
joy/distress, hope/f ear, admiration/reproach, pride/shame. Each pair can be represented by a variable whose value is in [−1, 1]. Thus, an emotion is rep4 resented by a vector V ∈ ([−1, 1]) . tions:
Following (DuyBui, 2004; Sehaba et al., 2007), three
Figure 1: Impacts of emotions on liking
dierent emotional vectors are dened in our model:
•
the emotions triggered by an event (whose values depend on both the triggering event and the NPC's personality);
• •
1996), anger is associated with a high degree of dominance, and distress and fear to a low degree. Consequently, we can model that emotion of pride or reproach
the emotional state of the NPC (which depends on
of
the entire aective experience of the character);
value that
2 the emotions expressed by the NPC . They can dier
tion of fear, distress, admiration, or shame of
from the emotional state since a character may decide to display an emotion dierent from its felt emotions (Prendinger and Ishizuka, 2001).
j
i
j
induces an increase in the dominance
thought to have on
j.
Conversely, an emo-
infers a decrease in the dominance
have on
i
i
caused
thought to
j.
In (Knutson, 1996), expression of sadness or fear reanger a high value. Finally, some types of emotions expressed by someone aect the dominance value of the
In the following, we analyze separately the dynamics of each variables of a social relation of a NPC
j
by
caused by
ects a low value of dominance and the expression of
Impacts of emotions on social relations
interaction with a character
i
i
during an
according to their emo-
person who perceives it. pression by
j
We can model that the ex-
of an emotion of fear or distress induces
an increase in the dominance of
i. Consequently, if the i and j with the same j will not change (Fig-
tions. If no event occurred, we suppose that the social
event triggers fear or distress for
relation does not evolve. In the contrary case, the in-
intensity the dominance of
tensities of the emotions triggered by the event are used
ure 2).
i
on
to update the social variables.
Liking.
(Ortony, 1991) shows that the degree of liking
one has for another depends on the emotions induced by the latter. Therefore, one's positive emotions triggered during an interaction with someone else may induce an increase in the degree of liking the former has for the latter. Conversely, negative emotions caused by another one may generate a decrease in the degree of liking (Ortony, 1991). However, an emotion triggered during an interaction with another person but not caused
Figure 2: Impacts of emotions on dominance
by the latter is supposed to have no impact on the degree of liking. For instance, let's imagine a situation in which Jack informs John about a very sad event, John will not necessarily like John less than before. Thus, we model that a positive emotion of
i
an increase in the degree of liking versely, negative emotion of
i
i
has
caused by
decrease in the degree of liking
i
j induces for j . Con-
caused by
has for
j generates a j (Figure 1).
Solidarity.
According
to
(de
Rivera
and
Grinkis,
1986), negative emotions caused by another person leads to a desire to reject him.
In other words, one's
negative emotion caused by someone else induces a decrease in the solidarity value that the former had with the latter. However, the triggering of positive emotions does not modify the solidarity value.
Indeed, the at-
traction resulting from a positive emotion corresponds According to (Shiota et al., 2004; Kelt-
to a desire to give to the other (such as tenderness) or
ner and Haidt, 2001), pride emotion reects dominance
to have something the other has (such envy) (de Rivera
whereas shame corresponds to an inferior status. Given
and Grinkis, 1986).
the denition of the combined emotion of anger in the
tion of
OCC model (Ortony et al., 1988), reproach appears
solidarity that
Dominance.
to reect dominance. In the PAD model (Mehrabian,
2
A character may express several emotions at the same time.
i
caused by
i
j
We model that a negative emoinduces a decrease in the value of
thought to have with
j
(Figure 3).
As highlighted in (Keltner and Haidt, 2001), during interpersonal relations, expressed emotions reect an individual's mental state (goals, beliefs, expectations, plan, etc).
Consequently, after the occurrence
of an event, a person perceiving another one's expression of emotion similar to her own triggered emotion, may imagine that they share similar goals or values. Conversely, opposite types of emotion may reect conicting goals or values. For instance, if an event triggers joy for a person because he had the goal that this event occurred and if, following this event, his interlocutor expresses distress, he may think that the latter had the opposite goal.
Finally, we model that the congruence
between one's triggered emotion and the emotion expressed by his interlocutor will inuence the solidarity variable. An incongruence (resp. congruence) may lead
Implementation
The
of
social
relations
presented
Let
Eti (evt)
above
is
In order to make it operational,
we have dened the update function
ϕ described below.
be the vector representing the intensities
i triggered by an event Exj (t) the vector of the intensities of the emotions expressed by the character j at time t. • We rst dene the function f : of the emotions of the character
evt
occurred at time
t,
and
f : (Eti (evt), Exj (t)) → ∆SocialRelationi,j (t)
to a decrease (resp. increase) in solidarity. If the triggered emotion of
model
semi-quantitative.
i is joy or hope (and is not caused by j )
This function takes as input the NPC's triggered
j expressed emotion of the same type, the solidarity increases. On the contrary, if j expresses an opposite and
type of emotion, the solidarity decreases (Figure 3).
emotions
and
tor,
returns
and
those a
expressed vector
by
its
representing
interlocuthe
vari-
ation of the social relation given these emotions (∆SocialRelationi,j (t)).
f is increasing for each ∆SocialRelationi,j (t) with respect to each dimension of the emotional vectors Eti (evt) and Exj (t) (except for the solidarity which increases dimension of
when the emotion values are close).
This function
returns the null-vector when the emotion vectors are null.
•
We then dene the function
g:
g : (social_relationi,j (t), ∆SocialRelationi,j (t)) → social_relationi,j (t + 1) This function takes as input the current social relation
social_relationi,j (t)
interlocutor
and
the
of
vector
the of
NPC
the
with
social
its
rela-
tion's variation returned by the previous function Figure 3: Impacts of emotions on solidarity
(∆SocialRelationi,j (t)). cial relation. and
Moreover, research shows that solidarity inuences liking.
i
pose the an increase in solidarity of increase in the degree of liking
i
with
has for
j
induces an
g
f
returns the updated so-
is increasing and has a low slope on
1
to represent the fact that the social relation
is hard to alter when on the extremes. For instance,
Indeed, it appears that one likes more similar
persons (Smith and Mackie, 2007). Then, we can sup-
−1
g
a sinus-based function can be used.
•
Finally, the function
j.
ϕ
is dened as follows:
ϕ : (social_relationi,j (t), Eti (evt), Exj (t)) → social_relationi,j (t + 1) This function takes as input three vectors represent-
F amiliarity.
In the literature, emotions seem to not
have a direct impact on the familiarity (i.e.
ing respectively the current social relation of the
on the
NPC with its interlocutor, its triggered emotions, and
degree of condentiality of the information transmitted
those expressed by its interlocutor. It returns the up-
by a person).
dated social relation.
However,
research shows that one
condes more in another when the former likes the latter (Collins and Miller, 1994). Therefore, we model that the more the character likes another one the more it will transfer a condential information to it. other words,
In
the familiarity is indirectly connected
to the liking variable.
Indeed,
an increase in the
condentiality of the information transferred leads to an increase in the familiarity value.
ϕ(social_relationi,j (t), Eti (evt), Exj (t)) = g(social_relationi,j (t), f (Eti (evt), Exj (t))) Note that in our model, the triggered and expressed emotions have to be provided.
Methods to compute
triggered emotions and to update a virtual character's emotional state consistently with its personality have been proposed in previous work (Prendinger et al., 2002; Sehaba et al., 2007). Dierent existing tools could
In the next section, we present an implementation of
be used to identify the emotions expressed by the char-
the model introduced above and illustrates it through
acter's interlocutor (both human and virtual) (Picard,
an example of interaction between two characters.
1997).
Illustrative example.
To illustrate the dynamics of a
triggers hope (here, the burglar would appreciate both
NPC's social relation, we have implemented the model
to face a comprehensive policeman and to drink a cof-
and tested it with the scene described in Table 1. The
fee).
context of the scene is a police interrogation (a burglar
the probability of the event and its desirability. For in-
facing a policeman) at the police station. After a break-
stance, in the fourth sentence, receiving a cup of coee
in in a jewelery, the police arrested the burglar but the
only triggers joy with a low intensity (0.2) whereas the
loot is missing.
8th sentence triggers both joy and hope with higher in-
The policeman wants the burglar to
confess where the money and the jewels are hidden.
The intensity of the emotion depends on both
tensities. On the contrary, the recall of an undesirable past event triggers distress (e.g. in the 7th sentence, the
0
Dialog between the Burglar s policeman and the Burglar triggered emotions 1.P oliceman: The facts are
not in your favor, you know. Ten people saw you threaten the manager with a weapon 2.Burglar: So, what do you want? 3.P oliceman: You know, I'm not a bad guy...(the policeman is preparing a cup of coffee) 4.P oliceman: Do you want some coee? (the policeman oers a cup of coee to the Burglar) 5.P oliceman: ...I know that your child has been kidnapped. Same thing happened to me last year, I had to negotiate with those b*** I know what it is (expression of distress) 6.P oliceman: I want to help
you. Just tell me where you hid the money 7.Burglar: I need this money to save my child! 8.P oliceman: We have received new information about the kidnappers. We know where your kid is being kept 9.P oliceman: Tell me everything and I'm sure I can nd a solution to avoid you going to jail 10.Burglar: Ok, I'm going to tell you what really happened...
Fear=0.9
policeman speaks about the kidnapping of the burglar's child). At the beginning of the interaction,
we suppose
that the burglar is submissive related to the policeman (dominance
= −0.3)
and dislikes him (liking
= −0.5).
Figure 4 illustrates the dynamics of the burglar's social relation with the policeman.
Hope=0.2
Joy=0.2
Distress=0.7
Hope=0.6
Figure 4: Dynamics of the burglar's social relation with the policeman Let us consider for instance the third utterance (hope with intensity 0.2).
Joy=0.6, Hope=0.8
As presented in the previous sec-
tion (see Figure 1), the emotion of hope triggered by the policeman increases the degree of liking the burglar has for the policeman. Since the intensity of the emotion is rather low and since it is the only positive emotion in
Hope=0.8
Table 1: Scene between a burglar and a policeman
this event, the impact will however stay low and no signicative change will happen. On the contrary, on the 5th interaction (distress with high intensity), since the policeman is not responsible for the negative emotion (he is not the kidnapper), the event has no impact on the degree of liking. Moreover, we have a congruence of the triggered emotion and the emotion expressed by the policeman.
As shown on Figure 3, this increases
the solidarity and, by side eect, the degree of liking. In this example, the values of social relation will
The left column in Table 1 show the dialogue between the policeman and the burglar. For each action or utterance of the policeman, the right column gives the triggered emotion of the burglar and its intensity, whose computation is based on the OCC model (Ortony et al., 1988).
For instance, let us consider the third
be used to determine which information the character transfers to its interlocutor. Since the degree of liking, dominance and familiarity reach given thresholds, the burglar will nally confess where the money and the jewel are.
sentence when the policeman reassures the burglar and
Conclusion
prepares some coee. According to OCC, the prospect
In this paper, we have presented a model of the dy-
of an event that would increase the NPC's satisfaction
namics of NPC's social relations taking as inputs the
emotions triggered by the scenario and other charac-
Fall Symposium on Socially Intelligent Agents - The
ters' actions.
Human in the Loop, North Falmouth.
Our approach takes into account vari-
ous results from cognitive science and psychology, and is then implemented in a social engine that computes and updates automatically the numeric variables describing the status of social relations. One motivation behind our design choices has been to limit the complexity (number of parameters, etc.) of our engine so that, though richer than most tools currently used by the game industry, it remains usable by game designers who are not trained in psychology and work on tight development schedules. Though we have only illustrated our work on a small, short dialog corresponding to a single scene in a game, we believe that this model and the corresponding social engine as a wide applicability for video games where the social credibility of NPCs actions is key (this includes notably role-playing games and adventure games). In
Isbister, K. (2006).
Better Game Characters by De-
sign: A Psychological Approach (The Morgan Kaufmann Series in Interactive 3D Technology).
Morgan
Kaufmann Publishers Inc., San Francisco, CA, USA. Keltner, D. and Haidt, J. (2001). Social functions of emotions. In Mayne, T. J. and Bonanno, G. A., editors, Emotions: Current issues and Future Directions, pages 192213. The Guilford Press. Knutson, B. (1996). Facial expressions of emotion inuence interpersonal trait inferences. Journal of Nonverbal Behavior, 20:165182. Mehrabian, A. (1996). Pleasure-arousal-dominance: A general framework for describing and measuring individual dierences in temperament. chology:
Current Psy-
Developmental, Learning, Personality, So-
the short term, we plan to incorporate this model in
cial, 14:261292.
game AI engine, currently being developed in collab-
Ortony, A. (1991). Memories, thoughts, and emotions:
oration with industry partners, so as to enhance the
Essays in honor of George Mandler, chapter Value and
quality of dialogs with NPCs in adventure games.
emotion. Hillsdale, NJ : Erlbaum.
In
the long term, our model can serve as a support for in-
Ortony, A. (2002).
teractive storytelling. Although the example presented
agents believable. In Trappl, R., Petta, P., and Payr,
here is very simple, it contains a number of dramatic
S., editors, Emotions in Humans and Artifacts, pages
elements that are explicitely described by the game pro-
189211. MIT Press.
grammer as emotional events in our model. In previous work (Sehaba et al., 2007), we presented an emotional model and engine that computes and updates a character's emotional state taking into account the character's personality and scenario events. We believe that the combination of all these elements will provide game developers with a tool that will greatly help the design of credible NPC's in immersive environments.
On making believable emotional
Ortony, A., Clore, G., and Collins, A. (1988). Press, United Kingdom.
Picard, R. (1997). Aective Computing. MIT Press. Prendinger, (2002).
H.,
Descamps,
S.,
and
Ishizuka,
M.
Scripting aective communication with life-
like characters in web-based interaction systems. Applied Articial Intelligence Journal, 16(7-8):519553. Prendinger, H. and Ishizuka, M. (2001).
Social role
awareness in animated agents. In AAMAS, pages 270
References
277, Montréal, Canada. ACM press.
André, E., Rehm, M., Minker, W., and Bühler, D. (2004). Endowing spoken language dialogue systems with emotional intelligence. In Aective Dialogue Systems, pages 178187. Bickmore, T. W. and Cassell, J. (2001).
The
cognitive structure of emotions. Cambridge University
Rousseau, D. and Hayes-Roth, B. (1998). psychological model for synthetic actors.
A socialIn Sycara,
K. P. and Wooldridge, M., editors, International Conference on Autonomous Agents, pages 165172, New
Relational
York. ACM Press.
agents: a model and implementation of building user
Sehaba, K., Sabouret, N., and Corruble, C. (2007). An
trust. In CHI, pages 396403.
emotional model for synthetic characters with person-
Brown, P. and Levinson, S. (1987). Politeness: Some
ality. In ACII Conference.
Universals in Language Usage. Cambridge University
Shiota, M., Campos, B., Keltner, D., and Hertenstein,
Press.
M. (2004). The Regulation of Emotion, chapter Posi-
Collins, N. and Miller, L. (1994). Self-disclosure and liking: A meta-analytic review. Psychological Bulletin, 116:457475. de Rivera, J. and Grinkis, C. (1986). Emotion as relationships. Motivation and Emotion, 10:351369. DuyBui, T. (2004). Creating emotions and facial expressions for embodied agents. PhD thesis, University of Twente. Gratch, J. (2000). Socially situated planning. In AAAI
tive Emotion and the Regulation of Interpersonal Relationships, pages 127155. Lawrence Erlbaum Associates. Smith, E. and Mackie, D. (2007). Social Psychology, chapter Liking and Loving, pages 393436. Psychology Press. Walker, M. A., Cahn, J. E., and Whittaker, S. J. (1997). Improvising linguistic style: Social and aective bases for agent personality. In First international conference on Autonomous Agents, pages 96105.